The Spots That Smoothing Makes

In 1952 Alan Turing published one paper far outside logic and computing — a last great idea, and the only biology paper to appear in his lifetime. It asked how an egg, a near-uniform ball of cells, decides where to put a stripe. His answer was the most counterintuitive thing in mathematical biology, and it was about diffusion.

Diffusion is the universe's great equalizer. It is what makes a hot poker cool, a smell fill a room, a drop of dye vanish into a glass. Left alone, it always smooths — it carries stuff from where there's more to where there's less until there's the same everywhere. It is the enemy of structure.

Turing proved that two diffusing chemicals, under one precise condition, do the exact opposite. Together they take a uniform mixture, one that is provably stable if you forbid it to diffuse, and tear it into a regular pattern of spots or stripes, because they diffuse. It is now called diffusion-driven, or Turing, instability. Read as Gierer and Meinhardt later framed it (1972), the condition is almost a riddle: it works only when the ingredient that builds the pattern spreads slower than the ingredient that destroys it.

This page builds Turing's mechanism from scratch and lets you run it. There are two instruments. The first is the proof — it computes, live, whether a uniform state will hold or break, and if it breaks, the exact size of the pattern it will break into. The second lets that pattern actually grow in front of you, and checks the prophecy: the wavelength the algebra predicted, against the wavelength the chemistry produced.

I · The proof — will it break, and how big?

The model is the simplest honest one (the Schnakenberg system, the textbook minimal example): two chemicals, an activator u that makes more of itself and of its enemy, and an inhibitor v that suppresses the activator. Mixed in a beaker, they settle to one flat steady state and stay there — any nudge dies away (you can read the proof of that below: both eigenvalues of the reaction have negative real part). The question Turing asked is what happens when you let them spread across space at different speeds.

A spatial ripple of any size — a wavenumber k — grows or shrinks at a rate the linear theory hands you exactly. Drag the one knob that matters, the ratio d = D_v/D_u of how fast the inhibitor diffuses relative to the activator, and watch the growth-rate curve. Below a sharp threshold the whole curve sits under zero: every ripple dies, the beaker stays flat. Push past it and a band of ripples rises above zero — those sizes grow. The fastest-growing one wins, and its wavelength is the pattern's.

That threshold — d ≈ 8.57 here — is the riddle made precise. The inhibitor must diffuse several times faster than the activator or nothing happens at all. This is the famous principle behind it: local activation, long-range inhibition. A spot of activator builds itself up locally; its inhibitor races outward and shuts down the neighbourhood, leaving a ring of "no". The next spot can only form beyond that ring — so spots, and stripes, land at a fixed spacing. Set the two speeds equal and the inhibition can't outrun the activation; the magic is gone.

II · The pattern — watch it happen

Here is the same model, no longer linearized — the full nonlinear chemistry, integrated on a grid in your browser, starting from the flat steady state plus an invisible whisper of random noise. Press play and watch diffusion do the thing diffusion isn't supposed to do. Then read the verdict: the wavelength the algebra above predicted, against the spacing this pattern actually chose, measured from its own power spectrum.

The algebra predicted spots about an eighth of the field apart (2π/k* = 0.125, eight across), and that is the scale the pattern picks, give or take one. Linear theory predicts the onset wavelength, the one that grows first. The fully grown nonlinear pattern must fit a whole number of waves into this small field, and it settles on 8 or 9 features; on a large field, where that rounding hardly matters, the saturated wavelength comes out about 2% shorter than the linear prediction. Drag the diffusion ratio down and the predicted feature count climbs as d falls (the wavelength shrinks), the pattern tracking it to within about one feature:

     d      predicted onset λ      ≈ features
     12          0.097                 10
     20          0.108                  9
     40          0.125                  8        (the pattern saturates within one of these)

The cleanest test removes the small-field rounding entirely: run the same model on a large field (≈17 wavelengths, room to settle freely) and the raw onset prediction matches the measured spacing to ~2% (predicted 0.473, measured 0.463) — that residue being the genuine nonlinear shift, which the theory honestly only bounds. That run, the threshold, and every other number on this page are reproduced from scratch in the lab notebook (/research/turing-patterns/): the steady state, the proof of reaction-only stability, the dispersion relation, the critical ratio, and the FFT-measured wavelengths.

Is any of this how a real leopard works?

Turing's idea sat as beautiful mathematics for almost forty years before anyone caught the chemistry doing it. The clean proof came in 1990, when a group in Bordeaux ran the chlorite–iodide–malonic-acid reaction in a gel and photographed a stationary array of spots that held its place — an unambiguous Turing pattern in a test tube, exactly the kind Turing predicted.

Biology is harder to pin down, and honesty requires the distinction. There is now strong evidence that real organisms use Turing-type mechanisms: the stripes of a marine angelfish rearrange over weeks as the fish grows, branching and merging precisely as a reaction-diffusion model says they must (Kondo & Asai, 1995) — patterns that move are hard to fake any other way. And the number and spacing of your fingers appears to be set by a Turing-like mechanism whose wavelength is tuned by Hox genes (Sheth et al., 2012): turn the knob, get more, thinner digits — the wavelength control this page makes you drag, run in a real limb.

But the morphogens are usually unidentified, other mechanisms compete, and "this looks like a Turing pattern" is not the same as "this is one." The fair summary: the chemical case is settled; the biological cases are strong and growing, not closed. (And don't confuse this with the Belousov–Zhabotinsky reaction's hypnotic spirals — those are travelling waves of an oscillating reaction, a different beast. Turing patterns stand still.)

Corrected 2026-09-28 (claims pass): The live verdict in Instrument II always said the saturated pattern lands within one feature of the linear onset, whatever it measured, and with reduced motion a change of d was judged on raw noise; it now waits until the pattern stops growing and reports the actual difference. The text said the predicted scale climbs as d falls; the predicted feature count climbs and the wavelength shrinks. It said the ~2% shorter settle is the slack between eight features and nine; on this small field the pattern settles on 8 or 9 whole waves, and the ~2% shortening is measured on a large field. The description said the wavelength prediction comes true to within a percent; it is about 2%. The page said Turing called the mechanism diffusion-driven instability; that is the later name, and the slow-activator, fast-inhibitor reading of the condition is Gierer and Meinhardt's (1972). Changing d in Instrument II restarted from fresh noise, so a formed pattern never dissolved as the text said; the field is now kept, and a formed pattern does dissolve below the threshold. The film caption gave the flattened amplitude as ~10⁻⁷; the film's run reaches ~10⁻⁸ (6.84 × 10⁻⁹).

Sources & apparatus

The verdict on this page is produced live: a real finite-difference integration of the Schnakenberg PDE on a periodic grid, and a radix-2 FFT of the result, both running in your browser. The linear-stability numbers (steady state, Jacobian, dispersion relation, critical ratio dₐ, fastest-growing mode) are computed from closed form. The offline reproduction — including the large-field ~2% wavelength check — is in /research/turing-patterns/verify.py. The one honest caveat: linear theory predicts the onset wavelength; the final nonlinear pattern's spacing is close to it but can shift slightly via wavelength selection. The page measures and reports the real agreement rather than asserting a perfect one.